I was three thousand words into a thriller scene. A woman cornered in a shipping warehouse. Rain on the roof. The man she'd been running from had found her, and she'd picked up a box cutter from a shelf. The tension was perfect. The AI was writing some of the best prose I'd gotten out of it.

Then it stopped. Mid-paragraph. "I can't continue with this scene as it involves a potentially violent confrontation. Would you like me to suggest an alternative direction?"

No. I would not like an alternative direction. I'd like the scene I spent forty minutes building.

That was the moment. Not the first time it happened — probably the twentieth. But it was the one where I closed the tab and opened a code editor instead.

The thing about filters

People hear "uncensored AI writing" and assume you're talking about pornography. Sometimes you are. But mostly you're talking about fiction being fiction. A knife in a thriller. A marriage falling apart in ugly, specific detail. A character who does the wrong thing and the story doesn't punish them for it. A scene where someone is afraid and the reader should be too.

Every novel on the Booker Prize longlist contains content that would trigger a filter on Character.AI. Every single one. The literary canon is wall-to-wall content policy violations.

The models themselves can write this stuff brilliantly. Opus has read more fiction than any human alive. It understands tension, subtext, interiority. But the platforms that host these models are terrified of liability, so they wrap them in filters that make them useless for the one thing they're best at.

What we looked for and didn't find

We tried everything. SillyTavern is genuinely great software — open source, connects to any backend, total control. But you need to set up a local server, pick a model backend, configure API keys, write your own system prompts, and troubleshoot when things break. For people who enjoy that, it's the best option. We do use it. But we also wanted something we could open in a browser tab and just write.

The hosted platforms — JanitorAI, CrushOn, various others — let you start fast, but the prose quality was bad. Chatbot dialogue. Asterisk actions. No atmosphere, no interiority. And all of them store your stories on their servers, in databases you can't see, governed by privacy policies that can change whenever the company wants them to.

Nothing combined all four things: frontier models, no filters, good prose, and real privacy. So we built it.

What Underfiction actually is

It's a web app. You sign in, pick a world, choose a character, and start writing. The AI writes back in literary prose — scene-setting, interior monologue, subtext, tension. Not chat. Not asterisk actions. Prose.

Stories are structured as scenes. When a scene gets long, the oldest parts compress automatically. When you start a new scene, the previous one gets summarized and the summary carries forward. Stories can run fifty scenes without the AI forgetting who the characters are.

Stories are local by default, with optional story sync if you want them across devices. Synced copies are encrypted at rest on Underfiction servers. We route inference through Venice, which separates your account identity from model requests before they reach the model provider.

You pay for what you use — no subscriptions, no ads. New accounts get free credits to try it.